Datasets:

voidful
/

agent-sft-stitch-zh-tts-taste-codec-chat-sample

conversation_id stringlengths 30 51	source stringclasses 16 values	speaker stringclasses 1 value	messages listlengths 5 51	rendered_chat stringlengths 6.22k 293k	input_ids listlengths 1.07k 75.9k	labels listlengths 1.07k 75.9k	length int32 1.07k 75.9k	message_count int32 5 51	assistant_turn_count int32 2 25	tool_result_turn_count int32 1 24	speech_segment_count int32 3 26	text_frame_count int32 40 673	codec_token_count int32 640 10.8k	supervised_token_count int32 812 53.1k	speech_segments listlengths 3 26	tokenizer_base stringclasses 1 value	expanded_vocab_size int32 367k 367k	format stringclasses 1 value
hermes_reasoning_default_0e5af74fd68212ee	hermes_reasoning	hung_yi_lee	[ { "role": "system", "content": "你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作流。第一個助理回覆必須立刻以 <SAY> 開頭。只有 <SAY>...</SAY> 內是合成語音，每個文字 token 後依序輸出 a 到 p 十六組 Taste-S FSQ code。不使用 Gemma 原生 audio tags。[SOPR]...[EOPR] 是模型的內部 reasoning，工具呼叫使用 <TOOL_CALL> 標記。以 <TOOL_RESULT> 標記的 user turn 是環境提供的工具結果，不可由助理自行生成。\n\n可用工具：\n[{\"name\...	<bos><\|turn>system 你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作流。第一個助理回覆必須立刻以 <SAY> 開頭。只有 <SAY>...</SAY> 內是合成語音，每個文字 token 後依序輸出 a 到 p 十六組 Taste-S FSQ code。不使用 Gemma 原生 audio tags。[SOPR]...[EOPR] 是模型的內部 reasoning，工具呼叫使用 <TOOL_CALL> 標記。以 <TOOL_RESULT> 標記的 user turn 是環境提供的工具結果，不可由助理自行生成。可用工具： [{"name": "whois", "description": "Fe...	[ 2, 105, 9731, 107, 237408, 213062, 44953, 240200, 238952, 55041, 238582, 238103, 180314, 236900, 5938, 4439, 54762, 236772, 236773, 50920, 240102, 30398, 11700, 237818, 236924, 141246, 180314, 237535, 240265, 70464, 100988, 237162, 655, 236773, 579...	[ -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100...	4,193	7	3	2	4	169	2,704	3,178	[ { "seg_index": 0, "text": "沒問題，我會為您查詢 amazon.com 的詳細 SSL 憑證資訊，包括核發者和有效期等細節。", "text_token_count": 30, "codec_token_count": 480, "assistant_content_token_count": 510 }, { "seg_index": 1, "text": "我正在確認 amazon.com 的網域基本資料，請稍等我一下。", "text_token_count": 18, "codec_token_count": 2...	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_aa010a56df4dd31e	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	4,890	7	3	2	4	226	3,616	4,119	[{"seg_index":0,"text":"沒問題，我很樂意為您的團隊找一些勵志名言。我會先從(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_c22b706426ca92f2	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	3,633	7	3	2	4	161	2,576	2,997	[{"seg_index":0,"text":"沒問題，面對這麼多選擇確實很難決定。我會先幫您查看(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_559ca072f4c191cd	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	6,341	11	5	4	6	276	4,416	5,248	[{"seg_index":0,"text":"沒問題，我明白您需要為 InnovaTech Inc. 建立一套完整的自(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_a260d6dead036c0c	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	2,699	7	3	2	4	123	1,968	2,351	[{"seg_index":0,"text":"沒問題，我會先幫您調取機器人 42 的感測器數據，接著再(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_24ab0caee2a91a65	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	4,304	5	2	1	3	205	3,280	3,740	[{"seg_index":0,"text":"沒問題，我會幫您搜尋 Documents 資料夾及其子目錄下的所(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_b6af90e0b621addd	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	2,494	7	3	2	4	112	1,792	2,147	[{"seg_index":0,"text":"沒問題，我會先幫您查看機器人 42 的感測器數據，接著再(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_a0615b982d963adc	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	3,928	7	3	2	4	176	2,816	3,290	[{"seg_index":0,"text":"沒問題，我明白面對這麼多好劇很難抉擇。我現在就來幫(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_8576bee1b02132c3	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	4,138	7	3	2	4	181	2,896	3,392	[{"seg_index":0,"text":"沒問題，我會立刻幫您處理。我會先更新學號 S1001 的學生(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3
hermes_reasoning_default_815a797ad4f980e0	hermes_reasoning	hung_yi_lee	[{"role":"system","content":"你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工(...TRUNCATED)	"<bos><\|turn>system\n你是一位台灣繁體中文語音助理，使用 STITCH-S 多輪工具工作(...TRUNCATED)	[2,105,9731,107,237408,213062,44953,240200,238952,55041,238582,238103,180314,236900,5938,4439,54762,(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	2,823	7	3	2	4	127	2,032	2,419	[{"seg_index":0,"text":"沒問題，我會用雞肉、番茄和蒜頭這幾樣食材，幫您搜尋(...TRUNCATED)	google/gemma-4-E2B	367,120	taste_s_fsq_interleaved_multiturn_v3

End of preview. Expand in Data Studio

Gemma 4 E2B Taste-S multi-turn codec SFT

This dataset contains 37,362 complete Traditional Chinese agent dialogues selected from voidful/agent-sft-stitch-zh-tts. It covers 229,434 synthesized speech segments, approximately 520.5 hours of audio before codec extraction.

Every assistant speech segment is represented without Gemma native audio tags:

<SAY> text_token <a_code> <b_code> ... <p_code> ... </SAY>

The first assistant output starts immediately with <SAY>. [SOPR]...[EOPR] is model reasoning, <TOOL_CALL> is supervised model output, and each <TOOL_RESULT> is an environment-provided user turn. System, original user, and tool-result tokens are masked with -100; all assistant outputs and turn endings are supervised.

Configurations

default: full messages, rendered chat, tensors, provenance, and statistics.
sft: only input_ids, labels, and length, ready for causal-LM collation.

from datasets import load_dataset

train_data = load_dataset(
    "voidful/agent-sft-stitch-zh-tts-taste-codec-chat-sample",
    "sft",
    split="train",
)

Statistics

Dialogues: 37,362
Assistant turns: 192,072
Assistant turns per dialogue: 2 to 25
Speech segments: 229,434
Text-aligned frames: 8,030,349
Codec tokens: 128,485,584
Sequence length: 1,066 to 75,852
Codec tokens per text frame: 16

Codec vocabulary

Gemma 4 E2B base vocabulary: 262,144
Added special tokens: 16 x 6,561 = 104,976
Expanded vocabulary: 367,120
ID formula: 262144 + codebook_index * 6561 + code
Token range: <a_0> (262144) through <p_6560> (367119)

The expanded model rows are newly initialized and require fine-tuning on this dataset before codec generation is meaningful. The matching local checkpoint uses a 131,072-token context window.

Downloads last month: 99